## ffmpeg / libavcodec / arm / dsputil_vfp.S @ 015f9f1a

History | View | Annotate | Download (6.93 KB)

1 |
/* |
---|---|

2 |
* Copyright (c) 2008 Siarhei Siamashka <ssvb@users.sourceforge.net> |

3 |
* |

4 |
* This file is part of FFmpeg. |

5 |
* |

6 |
* FFmpeg is free software; you can redistribute it and/or |

7 |
* modify it under the terms of the GNU Lesser General Public |

8 |
* License as published by the Free Software Foundation; either |

9 |
* version 2.1 of the License, or (at your option) any later version. |

10 |
* |

11 |
* FFmpeg is distributed in the hope that it will be useful, |

12 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |

13 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |

14 |
* Lesser General Public License for more details. |

15 |
* |

16 |
* You should have received a copy of the GNU Lesser General Public |

17 |
* License along with FFmpeg; if not, write to the Free Software |

18 |
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |

19 |
*/ |

20 | |

21 |
#include "config.h" |

22 |
#include "asm.S" |

23 | |

24 |
.syntax unified |

25 |
/* |

26 |
* VFP is a floating point coprocessor used in some ARM cores. VFP11 has 1 cycle |

27 |
* throughput for almost all the instructions (except for double precision |

28 |
* arithmetics), but rather high latency. Latency is 4 cycles for loads and 8 cycles |

29 |
* for arithmetic operations. Scheduling code to avoid pipeline stalls is very |

30 |
* important for performance. One more interesting feature is that VFP has |

31 |
* independent load/store and arithmetics pipelines, so it is possible to make |

32 |
* them work simultaneously and get more than 1 operation per cycle. Load/store |

33 |
* pipeline can process 2 single precision floating point values per cycle and |

34 |
* supports bulk loads and stores for large sets of registers. Arithmetic operations |

35 |
* can be done on vectors, which allows to keep the arithmetics pipeline busy, |

36 |
* while the processor may issue and execute other instructions. Detailed |

37 |
* optimization manuals can be found at http://www.arm.com |

38 |
*/ |

39 | |

40 |
/** |

41 |
* ARM VFP optimized implementation of 'vector_fmul_c' function. |

42 |
* Assume that len is a positive number and is multiple of 8 |

43 |
*/ |

44 |
@ void ff_vector_fmul_vfp(float *dst, const float *src0, const float *src1, int len) |

45 |
function ff_vector_fmul_vfp, export=1 |

46 |
vpush {d8-d15} |

47 |
fmrx r12, fpscr |

48 |
orr r12, r12, #(3 << 16) /* set vector size to 4 */ |

49 |
fmxr fpscr, r12 |

50 | |

51 |
vldmia r1!, {s0-s3} |

52 |
vldmia r2!, {s8-s11} |

53 |
vldmia r1!, {s4-s7} |

54 |
vldmia r2!, {s12-s15} |

55 |
vmul.f32 s8, s0, s8 |

56 |
1: |

57 |
subs r3, r3, #16 |

58 |
vmul.f32 s12, s4, s12 |

59 |
vldmiage r1!, {s16-s19} |

60 |
vldmiage r2!, {s24-s27} |

61 |
vldmiage r1!, {s20-s23} |

62 |
vldmiage r2!, {s28-s31} |

63 |
vmulge.f32 s24, s16, s24 |

64 |
vstmia r0!, {s8-s11} |

65 |
vstmia r0!, {s12-s15} |

66 |
vmulge.f32 s28, s20, s28 |

67 |
vldmiagt r1!, {s0-s3} |

68 |
vldmiagt r2!, {s8-s11} |

69 |
vldmiagt r1!, {s4-s7} |

70 |
vldmiagt r2!, {s12-s15} |

71 |
vmulge.f32 s8, s0, s8 |

72 |
vstmiage r0!, {s24-s27} |

73 |
vstmiage r0!, {s28-s31} |

74 |
bgt 1b |

75 | |

76 |
bic r12, r12, #(7 << 16) /* set vector size back to 1 */ |

77 |
fmxr fpscr, r12 |

78 |
vpop {d8-d15} |

79 |
bx lr |

80 |
endfunc |

81 | |

82 |
/** |

83 |
* ARM VFP optimized implementation of 'vector_fmul_reverse_c' function. |

84 |
* Assume that len is a positive number and is multiple of 8 |

85 |
*/ |

86 |
@ void ff_vector_fmul_reverse_vfp(float *dst, const float *src0, |

87 |
@ const float *src1, int len) |

88 |
function ff_vector_fmul_reverse_vfp, export=1 |

89 |
vpush {d8-d15} |

90 |
add r2, r2, r3, lsl #2 |

91 |
vldmdb r2!, {s0-s3} |

92 |
vldmia r1!, {s8-s11} |

93 |
vldmdb r2!, {s4-s7} |

94 |
vldmia r1!, {s12-s15} |

95 |
vmul.f32 s8, s3, s8 |

96 |
vmul.f32 s9, s2, s9 |

97 |
vmul.f32 s10, s1, s10 |

98 |
vmul.f32 s11, s0, s11 |

99 |
1: |

100 |
subs r3, r3, #16 |

101 |
vldmdbge r2!, {s16-s19} |

102 |
vmul.f32 s12, s7, s12 |

103 |
vldmiage r1!, {s24-s27} |

104 |
vmul.f32 s13, s6, s13 |

105 |
vldmdbge r2!, {s20-s23} |

106 |
vmul.f32 s14, s5, s14 |

107 |
vldmiage r1!, {s28-s31} |

108 |
vmul.f32 s15, s4, s15 |

109 |
vmulge.f32 s24, s19, s24 |

110 |
vldmdbgt r2!, {s0-s3} |

111 |
vmulge.f32 s25, s18, s25 |

112 |
vstmia r0!, {s8-s13} |

113 |
vmulge.f32 s26, s17, s26 |

114 |
vldmiagt r1!, {s8-s11} |

115 |
vmulge.f32 s27, s16, s27 |

116 |
vmulge.f32 s28, s23, s28 |

117 |
vldmdbgt r2!, {s4-s7} |

118 |
vmulge.f32 s29, s22, s29 |

119 |
vstmia r0!, {s14-s15} |

120 |
vmulge.f32 s30, s21, s30 |

121 |
vmulge.f32 s31, s20, s31 |

122 |
vmulge.f32 s8, s3, s8 |

123 |
vldmiagt r1!, {s12-s15} |

124 |
vmulge.f32 s9, s2, s9 |

125 |
vmulge.f32 s10, s1, s10 |

126 |
vstmiage r0!, {s24-s27} |

127 |
vmulge.f32 s11, s0, s11 |

128 |
vstmiage r0!, {s28-s31} |

129 |
bgt 1b |

130 | |

131 |
vpop {d8-d15} |

132 |
bx lr |

133 |
endfunc |

134 | |

135 |
#if HAVE_ARMV6 |

136 |
/** |

137 |
* ARM VFP optimized float to int16 conversion. |

138 |
* Assume that len is a positive number and is multiple of 8, destination |

139 |
* buffer is at least 4 bytes aligned (8 bytes alignment is better for |

140 |
* performance), little endian byte sex |

141 |
*/ |

142 |
@ void ff_float_to_int16_vfp(int16_t *dst, const float *src, int len) |

143 |
function ff_float_to_int16_vfp, export=1 |

144 |
push {r4-r8,lr} |

145 |
vpush {d8-d11} |

146 |
vldmia r1!, {s16-s23} |

147 |
vcvt.s32.f32 s0, s16 |

148 |
vcvt.s32.f32 s1, s17 |

149 |
vcvt.s32.f32 s2, s18 |

150 |
vcvt.s32.f32 s3, s19 |

151 |
vcvt.s32.f32 s4, s20 |

152 |
vcvt.s32.f32 s5, s21 |

153 |
vcvt.s32.f32 s6, s22 |

154 |
vcvt.s32.f32 s7, s23 |

155 |
1: |

156 |
subs r2, r2, #8 |

157 |
vmov r3, r4, s0, s1 |

158 |
vmov r5, r6, s2, s3 |

159 |
vmov r7, r8, s4, s5 |

160 |
vmov ip, lr, s6, s7 |

161 |
vldmiagt r1!, {s16-s23} |

162 |
ssat r4, #16, r4 |

163 |
ssat r3, #16, r3 |

164 |
ssat r6, #16, r6 |

165 |
ssat r5, #16, r5 |

166 |
pkhbt r3, r3, r4, lsl #16 |

167 |
pkhbt r4, r5, r6, lsl #16 |

168 |
vcvtgt.s32.f32 s0, s16 |

169 |
vcvtgt.s32.f32 s1, s17 |

170 |
vcvtgt.s32.f32 s2, s18 |

171 |
vcvtgt.s32.f32 s3, s19 |

172 |
vcvtgt.s32.f32 s4, s20 |

173 |
vcvtgt.s32.f32 s5, s21 |

174 |
vcvtgt.s32.f32 s6, s22 |

175 |
vcvtgt.s32.f32 s7, s23 |

176 |
ssat r8, #16, r8 |

177 |
ssat r7, #16, r7 |

178 |
ssat lr, #16, lr |

179 |
ssat ip, #16, ip |

180 |
pkhbt r5, r7, r8, lsl #16 |

181 |
pkhbt r6, ip, lr, lsl #16 |

182 |
stmia r0!, {r3-r6} |

183 |
bgt 1b |

184 | |

185 |
vpop {d8-d11} |

186 |
pop {r4-r8,pc} |

187 |
endfunc |

188 |
#endif |